While AI helps to automate complex decision-making tasks, it also reduces human agency. For some applications such as search, information filtering, writing and sensemaking, this can lead to catastrophic outcome like getting trapped in information bubbles and increased susceptibility to biased viewpoints. In this line of research, we investigate techniques that mitigate these harmful effects of AI powered information systems by supporting human agency in AI systems.

AI systems are often designed by individuals, teams, and organizations with a secondhand view of end-user workflow. This leads to end users often modifying their existing workflows to accommodate AI systems, which in many cases decreases productivity. Our research on transfer learning focuses on scaffolding the sensemaking process in AI system design, thus enabling end users with limited expertise in AI to inform the design of AI. And no, this is not prompt “engineering”.

When seeking machine assistance in complex and high-stakes decision-making tasks, users often need to understand why and how the machine is suggesting an action. Incorporating this transparency is critical for usable AI systems, and to help users calibrate their reliance on AI. This line of research focuses on incorporating transparency, accountability, and appropriate reliance within AI systems used in high-stakes settings.